Data-aware review
Schema-only review tells you a change looks risky. Data-aware review tells you whether it will actually fail on your data — and, on request, proves it. Opt-in, least-privilege, off by default.
Tier 1 — read-only statistics
Sieve checks the change against the planner's own statistics: row counts, null fractions, distinctness, and index coverage. That is enough to know a NOT NULL will abort because 12% of the column is already null, or that a blocking lock is brief at 800 rows but minutes at 40M.
- No row of real data is ever read — only the catalog numbers.
- The same numbers can be collected inside your own CI, so database credentials never leave your perimeter.
Tier 2 — shadow run
On request, Sieve applies the candidate migration to a throwaway copy of the database inside a transaction, measures the real outcome — duration, table rewrites, and actual constraint failures — and then rolls back. The dangerous migration fails on a disposable branch instead of in production.
- Non-transactional statements (e.g.
CREATE INDEX CONCURRENTLY) are detected and skipped, not faked. - Rewrites are caught by watching the table's underlying file change, not by guessing.